Storage tank life evaluation method based on test and flight data fusion
By integrating ground test and flight data, and employing principal component analysis, improved DS evidence theory, and LSTM models, the accuracy and reliability issues of single-aircraft life assessment were resolved, achieving high-precision life assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-03-24
AI Technical Summary
In existing single-aircraft life assessment methods, there are discrepancies between ground test data and actual flight data, resulting in inaccurate and unreliable assessment results, making it difficult to comprehensively cover the damage evolution patterns throughout the entire life cycle of the structure.
Principal component analysis was used to extract lifetime-related features. Combined with improved DS evidence theory and an attention-based LSTM model, ground test data and flight data were integrated. The model parameters were optimized through dynamic weight adjustment and Bayesian inference to achieve more accurate lifetime assessment.
It improves the accuracy and reliability of single-unit life assessment, significantly reduces assessment error, and controls the error within 5%, making it suitable for operation and maintenance decisions of aviation equipment.
Smart Images
Figure CN121723398A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aircraft structural life assessment technology, and more specifically, to a single-aircraft life assessment method based on the fusion of test and flight data. Background Technology Life assessment of aircraft structures is a core component of ensuring flight safety and optimizing operation and maintenance strategies, directly impacting the operational effectiveness and cost of aviation equipment. Currently, single-aircraft life assessment methods are mainly divided into two categories: one is based on ground test data, which establishes a life assessment model by simulating the structural stress and damage evolution process under typical operating conditions; the other is based on actual flight data, utilizing load, strain, and other information collected by onboard sensors to analyze the cumulative damage of the structure in real time.
[0002] However, existing methods have obvious limitations: when relying solely on ground test data, the test conditions differ from actual flight conditions, and the simulated environment cannot fully reproduce complex dynamic factors such as airflow and maneuvering loads, resulting in a large deviation between the evaluation results and the actual lifespan of a single aircraft; when relying solely on flight data, the integrity and reliability of the data are difficult to guarantee due to factors such as sensor accuracy, data transmission delay, and lack of extreme condition data, and a single data source cannot cover the damage evolution pattern of the entire life cycle of the structure, which easily leads to unstable evaluation results.
[0003] Therefore, how to effectively integrate systematic data from ground tests with personalized data from actual flights to compensate for the deficiencies of a single data source and improve the accuracy and reliability of single-aircraft life assessment has become a pressing technical challenge. Summary of the Invention
[0004] In view of this, the present invention provides a tank life assessment method based on the fusion of test and flight data, which can overcome the limitations of the single data source in the existing single-aircraft life assessment methods. It provides a single-aircraft life assessment method based on the fusion of test and flight data, and achieves more accurate and reliable life assessment results through the complementarity and synergy of multi-source data, providing a scientific basis for aircraft operation and maintenance decisions.
[0005] To achieve the above objectives, the technical solution of the present invention includes the following steps: Step (1) Collect ground test data and tank flight data, and perform noise reduction and standardization processing.
[0006] Step (2) uses principal component analysis to extract lifetime-related features.
[0007] Step (3) Construct a dual-source data fusion model based on the improved DS evidence theory, fuse the extracted core features, and obtain a comprehensive feature vector.
[0008] Step (4) Construct a lifetime assessment model based on fusion features, input the integrated feature vector to obtain the remaining lifetime assessment result of a single machine, and optimize the model parameters through cross-validation during the process.
[0009] Furthermore, in step (1), outliers are removed from the ground test data using the 3σ criterion, and discontinuous data are supplemented by cubic spline interpolation.
[0010] Furthermore, in step (1), the Kalman filter algorithm is used to eliminate sensor noise in the tank flight data, and invalid data is removed in combination with the flight phase judgment; the dual-source data are uniformly converted into the same time and physical quantity dimensions.
[0011] Further, in step (2), principal component analysis is used to extract lifetime-related features. Specifically, static features are extracted from ground test data; dynamic features are extracted from tank flight data; and principal component analysis is used to reduce the dimensionality of the extracted features, retaining core features with a contribution rate greater than 90%.
[0012] Furthermore, static features, including damage accumulation rate, low-temperature fatigue limit, and stiffness attenuation coefficient, are extracted from ground test data; dynamic features, including dynamic pressure peak value, pressure spectral density, and strain cycle number, are extracted from tank flight data.
[0013] Further, step (3) is specifically as follows: the initial weights of the dual-source data are determined by the analytic hierarchy process. When the integrity of the tank flight data is greater than 95%, the weight of the flight data is adjusted to 0.6-0.7, and vice versa. The core features are fused by the evidence synthesis rules to obtain the comprehensive feature vector.
[0014] Further, step (4) is specifically as follows: construct a long short-term memory network model based on the attention mechanism, input the comprehensive feature vector to obtain the evaluation result of the remaining launch cycles of the tank, which is the evaluation result of the remaining life of the single machine; model training uses the fusion sample of experimental data and flight data for supervised learning, and optimizes the model parameters through cross-validation.
[0015] Furthermore, step (4) also includes the following steps to correct the results of the single unit's remaining life assessment: combining the tank's historical maintenance records and post-launch damage assessment data, if the assessment results deviate from the historical damage patterns by more than 5%, the model parameters are updated based on Bayesian inference to correct the assessment results.
[0016] Beneficial effects: (1) This invention compensates for the shortcomings of a single data source by integrating ground test and single-aircraft flight dual-source data. Test data provides a systematic damage evolution law for evaluation, while flight data reflects the personalized operating conditions of a single aircraft. The dual-source complementarity improves the comprehensiveness and reliability of the data.
[0017] (2) This invention uses an improved DS evidence theory to achieve data fusion, and dynamically adjusts the weights based on data integrity, which solves the problem of fusion deviation caused by the difference in credibility between the two sources of data and improves the accuracy of feature fusion.
[0018] (3) The present invention adopts an LSTM model based on attention mechanism, which can capture the dynamic relationship between features and lifetime. Combined with Bayesian inference for result correction, the accuracy of lifetime assessment is further improved and the assessment error is significantly reduced compared with traditional methods. Attached Figure Description
[0019] Figure 1 These are the key steps of the tank life assessment method based on the fusion of test and flight data of this invention; Figure 2 This invention improves the structure of the DS evidence theory model; Figure 3 This invention presents the LSTM model structure based on the attention mechanism. Detailed Implementation
[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0021] This invention provides a single-aircraft life assessment method based on the fusion of test and flight data, comprising the following steps: (1) The collected ground test data and flight data were denoised and standardized respectively; for the ground test data, outliers caused by the error of the test equipment were removed and intermittent data were supplemented by data interpolation; for the single-aircraft flight data, the Kalman filter algorithm was used to eliminate sensor noise and invalid data under extreme conditions were removed in combination with the flight status judgment; then the dual-source data were uniformly converted into the same time and physical quantity dimensions to provide a basis for fusion analysis.
[0022] (2) Based on the preprocessed dual-source data, life-related features are extracted; static features such as damage accumulation rate, fatigue limit, and stiffness attenuation coefficient of the structure under different load levels are extracted from ground test data; dynamic features such as dynamic load peak, load spectral density, and strain cycle number are extracted from single-aircraft flight data; principal component analysis (PCA) is used to reduce the dimensionality of the extracted features, retain the core features with a contribution rate greater than 90%, and reduce data redundancy.
[0023] (3) Construct a dual-source data fusion model based on the improved DS evidence theory and fuse the extracted core features. First, determine the initial weights of ground test data and single-aircraft flight data by expert scoring method, and then dynamically adjust the weights based on data credibility. When the integrity of flight data is greater than 95%, increase the weight of flight data to 0.6-0.7, and vice versa. Then, fuse the dual-source feature information through evidence synthesis rules to obtain a comprehensive feature vector. Figure 2 This invention improves the structure of the DS evidence theory model.
[0024] (4) Construct a life assessment model based on fusion features, and input the comprehensive feature vector to obtain the remaining life assessment result of a single aircraft; the life assessment model is based on a long short-term memory network (LSTM), and an attention mechanism is introduced to strengthen the influence of key features on life assessment. During the model training process, fusion samples of experimental data and flight data are used for supervised learning, and the model parameters are optimized through cross-validation. Figure 3 This invention presents the LSTM model structure based on the attention mechanism.
[0025] (5) Combine the historical operation and maintenance records of the single machine with the structural inspection data to correct the life assessment results; if the assessment results deviate from the historical damage patterns by more than 5%, update the model parameters based on Bayesian inference to ensure that the assessment results are consistent with the actual state of the single machine.
[0026] More specifically: Taking the life assessment of a spacecraft liquid oxygen tank as an example, this invention provides a more detailed description of a single-unit life assessment method based on the fusion of experimental and flight data. Figure 1 The steps are as follows: Step (1): The collected weld crack length under stress impact is used as damage evolution data for preprocessing, where the failure criterion is that the crack length reaches 0.3 mm.
[0027] For ground test data, outliers in strain and internal pressure data (i.e., strain values exceeding the range of μ + 3σ) were removed using the 3σ criterion, with a removal rate of approximately 2.1%, where μ is the mean of each parameter and σ is the standard deviation of the parameter. For intermittent data caused by test interruptions, cubic spline interpolation was used to supplement the data to ensure continuity. A piecewise cubic polynomial was constructed to ensure that the interpolation function satisfies continuity, smoothness, and boundary conditions at given data points, i.e., within the interval... Above, cubic spline The expression is: (1) in, All of these are coefficients to be determined.
[0028] Using the continuity of the second derivative and boundary conditions, we obtain information about... The linear equations with second derivatives were used to obtain the coefficients and cubic polynomials for each interval. The second derivative of the natural spline is zero at the breakpoint, and the curve tends to flatten at the breakpoint. With an interpolation error range of ±1.8%, the processed data included time, internal pressure, strain, and crack length.
[0029] For the actual measurement data of the storage tank, the Kalman filter algorithm is used to eliminate sensor noise. First, a Kalman filter prediction model, i.e., the state prediction equation, is constructed: (2) in, Here is the state transition matrix. To control the input matrix, To control the input vector, This represents Gaussian white noise.
[0030] Construct the observation equation: (3) in, For the observation matrix, To observe noise.
[0031] By setting the filter coefficient to 0.85 and combining it with the flight phase judgment to remove invalid data under conditions such as launch escape and extreme aerodynamic interference, the signal-to-noise ratio of strain data was improved from 22dB to 45dB, with a removal rate of approximately 2.8%.
[0032] Step (II): Extract four static features from the preprocessed ground test data: damage accumulation rate, low-temperature fatigue limit, stiffness attenuation coefficient, and yield strength; extract four dynamic features from the tank flight data: peak pressure, pressure spectral density, strain cycle count, and temperature cycle amplitude; use principal component analysis (PCA) to reduce the dimensionality of these eight features. First, standardize the eight features, then calculate the covariance matrix of the data to obtain the correlation between the features. (4) Perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalues. and the corresponding feature vector eigenvalues This represents the variance contribution of the i-th principal component.
[0033] Select the eigenvectors corresponding to the first k largest eigenvalues. The projection matrix is formed, where k is 4, determined by the cumulative contribution rate of 93.1%, and finally the damage accumulation rate, dynamic pressure peak, strain cycle number, and low temperature fatigue limit are obtained.
[0034] Step (3): By introducing the Pearson correlation coefficient to optimize the weight allocation, an improved Dempster-Shafer (DS) evidence theory is proposed. Ground and flight data are input, and the four core features are fused through the DS evidence synthesis rules to give a comprehensive judgment after data fusion, obtain a comprehensive feature vector, and thus clearly express the degree of uncertainty.
[0035] First, the importance of the dual-source data was scored using an expert scoring method, and the initial weight of the ground test data was determined to be 0.5, while the initial weight of the tank flight data was 0.45. Then, the weights were dynamically adjusted based on the integrity of the flight data—the weight value of data with integrity >95% was adjusted to 0.62-0.68, and the weight of test data with integrity ≤95% was adjusted to 0.6-0.65.
[0036] When the integrity of the tank flight data is 97%, the weight of the flight data is adjusted to 0.62 and the weight of the test data is adjusted to 0.38. The scoring table is shown in Table 1.
[0037] Table 1 Expert Weighting Scoring Table
[0038] First, define the recognition framework. , which represents the set of all mutually exclusive and complete basic assumptions about the problem.
[0039] Calculate the conflict coefficient, i.e.: (5) in It indicates the degree of direct support of evidence for subset A, that is, the confidence assigned to propositions that are exactly A.
[0040] If the conflict coefficient K is too high, it may lead to unreliable results. Based on the conflict coefficient, the Dempster combination rule can be obtained: (6) Step (iv): Construct an LSTM lifetime assessment model based on an attention mechanism. Set the model input as a 4-dimensional comprehensive feature vector and the output as the remaining number of firing cycles in the tank. During model training, use 70% of the fused samples as the training set and 30% as the test set. Set the number of hidden layer nodes to 80, the learning rate to 0.0012, and the number of iterations to 120 rounds. Optimize the model parameters through cross-validation to ensure the model's generalization ability and adapt to the damage assessment requirements in low-temperature environments.
[0041] Processing input window We obtain the hidden states generated by the LSTM at each time step, forming a matrix H. The hidden states encode the device and degradation information at each local time step.
[0042] The hidden state at the last time step represents the current state. h t With all historical hidden states h i The matching degree is calculated, and the weight score is converted into an attention weight distribution, with the sum of each weight being 1.
[0043] All weights Apply to the corresponding hidden state Perform a weighted summation to obtain the context vector. , As a dynamic and focused information summary, it concentrates the historical information most relevant to the current prediction while downplaying irrelevant information.
[0044] Finally, the context vector The final hidden state of the LSTM is concatenated with the concatenated state, and the resulting string is used for regression calculation through a fully connected layer to obtain a scalar value for the remaining lifetime.
[0045] Step (5): Retrieve the historical operation and maintenance records of the storage tanks, including weld inspection reports, cryogenic sealing performance test data, replacement component information, and post-launch damage assessment data. Compare the life assessment results with historical damage patterns. For storage tanks with an assessment deviation exceeding 5%, use Bayesian inference to update the prior probability of the model. After correction, the assessment error is reduced to within 2.5%, ensuring that the assessment results accurately match the actual service status of the storage tanks. The final assessment results and errors for each storage tank are shown in Table 2.
[0046] Table 2 Lifetime Assessment Results
[0047] The results in the table above demonstrate that the proposed method achieves high accuracy, with the final life assessment error for all 12 tanks controlled within 5%, and an average error of 2.8%. In this embodiment, dual-source data fusion compensates for the difference between the "standardized low-temperature environment" of ground tests and the "dynamic temperature-pressure coupling" of actual flights. The improved DS evidence theory addresses the dynamic matching problem of the credibility of test data and flight data, while the attention mechanism LSTM accurately captures the nonlinear correlation between tank weld damage and characteristic parameters. These three elements work together to achieve high-precision assessment. This invention provides a reference template for the engineering application of life assessment for tanks and similar applications.
[0048] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for assessing tank life based on the fusion of test and flight data, characterized in that, Includes the following steps: Step (1) Collect ground test data and tank flight data, and perform noise reduction and standardization processing; Step (2) uses principal component analysis to extract lifetime-related features; Step (3) Construct a dual-source data fusion model based on the improved DS evidence theory, fuse the extracted core features, and obtain a comprehensive feature vector; Step (4) Construct a lifetime assessment model based on fusion features, input the integrated feature vector to obtain the remaining lifetime assessment result of a single machine, and optimize the model parameters through cross-validation during the process.
2. The tank life assessment method based on the fusion of test and flight data as described in claim 1, characterized in that, In step (1), outliers are removed from the ground test data using the 3σ criterion, and discontinuous data are supplemented by cubic spline interpolation.
3. The tank life assessment method based on the fusion of test and flight data as described in claim 1, characterized in that, In step (1), the Kalman filter algorithm is used to eliminate sensor noise in the tank flight data, and invalid data is removed in combination with the flight phase judgment; the dual-source data are uniformly converted into the same time and physical quantity dimensions.
4. The tank life assessment method based on the fusion of test and flight data as described in claim 1, characterized in that, Step (2) uses principal component analysis to extract lifetime-related features. Specifically, the method is to extract static features from ground test data. Dynamic features were extracted from the tank flight data; principal component analysis was used to reduce the dimensionality of the extracted features, retaining the core features with a contribution rate greater than 90%.
5. The tank life assessment method based on the fusion of test and flight data as described in claim 4, characterized in that, The static features extracted from the ground test data include damage accumulation rate, low-temperature fatigue limit, and stiffness attenuation coefficient; the dynamic features extracted from the tank flight data include dynamic pressure peak value, pressure spectral density, and strain cycle number.
6. The tank life assessment method based on the fusion of test and flight data as described in claim 1, characterized in that, The specific method of step (3) is as follows: the initial weight of the dual-source data is determined by the analytic hierarchy process. When the integrity of the tank flight data is greater than 95%, the weight of the flight data is adjusted to 0.6-0.7, and vice versa. The core features are fused by the evidence synthesis rules to obtain the comprehensive feature vector.
7. The tank life assessment method based on the fusion of test and flight data as described in claim 1, characterized in that, The specific method of step (4) is as follows: construct a long short-term memory network model based on attention mechanism, input the comprehensive feature vector to obtain the evaluation result of the remaining launch cycle number of the tank, which is the evaluation result of the remaining life of the single machine; model training adopts supervised learning using the fusion sample of experimental data and flight data, and optimizes the model parameters through cross-validation.
8. The tank life assessment method based on the fusion of test and flight data as described in claim 1, characterized in that, Step (4) further includes the following steps: correcting the results of the single unit remaining life assessment by combining the historical operation and maintenance records of the storage tank with the post-launch damage assessment data. If the assessment results deviate from the historical damage patterns by more than 5%, the model parameters are updated based on Bayesian inference to correct the assessment results.